Greater ARC will usually come with improved performance, though note that this is read caching, so you only ever need it to be as large as your “hot set”. On file servers serving TBs of data from slow drives without SSD caching, you may want tens of GBs of ARC.
But for most other cases where I/O to the underlying disk aren’t bad to start with and you don’t have GB/s of writes, having an ARC of just a few GBs is usually fine.
My understanding of recordsize is that it works in much the same way as a block size, so if you set it to a very high value, small writes will use at least one record and so potentially waste space. Using too small a value leads to needless work and fragmentation. The ideal value for something like a database would be to line it up with the size of writes performed by the database, so that when it writes something to disk, it fills exactly one record or more record.